Railway track systems constitute a complex coupled structural problem, positioned between the supporting infrastructure below—such as bridges, tunnels, and earthworks—and the rolling stock operating above the rails. Accurate analysis therefore requires explicit consideration of both the train–track and track–infrastructure interfaces. Traditional modelling approaches have treated vehicle dynamics, track dynamics, and structural dynamics as largely independent disciplines, relying on simplified boundary conditions such as rigid track support, static or harmonic vehicle loading, and rigid infrastructure assumptions. While these simplifications were acceptable for earlier railway systems with lower speeds and axle loads, they are no longer sufficient for modern high-speed and heavy-haul operations.

Increasing operational demands have amplified dynamic interaction effects, leading to critical issues including reduced running safety, excessive track settlement, rail corrugation, wheel defects, elevated noise levels, and compromised ride comfort. Consequently, the development of advanced coupled analysis models that accurately capture train–track–infrastructure interaction has become essential, yet existing models remain limited and insufficiently validated against real-time field measurements.

In parallel, significant advances in artificial neural networks and machine learning have introduced new methodologies for solving complex physical problems, particularly through the emerging field of Scientific Machine Learning. Physics-Informed Neural Networks represent a promising approach by embedding governing physical laws directly into the learning process, enabling data-efficient and physically consistent solutions to ordinary and partial differential equations. This thesis aims to integrate PINN methodology with a comprehensive dynamic model of railway track systems, incorporating interactions with both rolling stock and supporting infrastructure. By utilizing real-time measurement data obtained from vehicles, track components, and infrastructure alongside numerical simulations, the proposed framework seeks to deliver more accurate and reliable predictions of system behaviour.